Inspired by iSPIRT’s DEPA — the Data Empowerment and Protection Architecture.

Train AI Models on Confidential Data Without Compromising Privacy

A data provider, a model owner, and a clean-room provider sign one contract. Training then runs on the cluster that provider already operates. The raw dataset stays with the provider.

How a job runs

One contract, then one training run on a clean room the provider already operates.

1. Agree

The data provider publishes a dataset. The model owner writes a Ricardian contract that names the dataset, the use, and the clean-room provider.

2. Sign

Each party signs. Training does not start until the contract is signed.

3. Train

The job runs on that provider’s cluster. The result, the logs, and the provenance stay with the contract.

Use the cluster you already run

The contract names the cluster. The kubeconfig path stays on the provider’s Cloud Provider card.

Local

Training stays on this machine.

Azure

An existing AKS cluster. The job does not create a subscription.

Google Cloud

An existing GKE cluster. The job does not create a project.

Oracle Cloud

An existing OKE cluster. AWS does not start a training job.

A missing kubeconfig path fails the job. Attestation stays simulated until a SEV-SNP guest report verifies.

What the record shows

The same contract is the place to check signatures, training, and audit.

Signatures

Each party’s signature is stored on the contract before training starts.

Keys

A key release stores the time of release. The key bytes are not kept.

Audit

An auditor checks a sealed Merkle root in this system. That check is not a filing with an external records office.

Ask a question, or help build it

Write to hello@dpi-apps.space.